Creativity and Constraints: The Production of Novel Sentences
Catrinel Haught ([email protected])
Department of Psychology, Green Hall, Princeton University
Princeton, NJ 08544 USA
Philip N. Johnson-Laird ([email protected])
Department of Psychology, Green Hall, Princeton University
Princeton, NJ 08544 USA
Abstract
Two experiments explored how people create novel sentences
referring to given entities presented either in line drawings or
in nouns. The line drawings yielded more creative sentences
than the words, both as rated by judges and objectively in
Shannon’s measure of the amount of information that the
sentences conveyed. A hypothesis about the cognitive
processes of creation predicted this result: creativity depends
on constraints. Line drawings of entities present more
information about them than nouns denoting the same entities,
and so the pictures provide more constraints than the nouns.
Hence, line drawings yield more creative sentences than
words.
Introduction
Some of the most compelling instances of creativity occur in
the production of language, which relies on the power of
recursion (Chomsky, 1965). But how individuals generate
novel sentences remains a mystery. In particular, the
cognitive work leading to a thought, let alone a creative one,
and to its formulation in a well-formed sentence lacks an
adequate explanation (Levelt, 1989). The experiments
described below aimed to make progress in elucidating this
mystery and the deeper mystery of creativity. Their focus
was not on the generation of grammatically correct
sequences of words, but on how individuals can incorporate
given referents into a novel sentence.
This research follows in the “creative cognition” tradition
(Finke, Ward, & Smith, 1992) and its theoretical motivation
derives in part from the NONCE analysis of creativity
(Johnson-Laird, 2002). This definition of creativity
postulates that the outcome of a creative process is: Novel
for the person producing the result, Optionally novel for
society at large, the result of a Nondeterministic process that
is guided by Constraints and that is based on existing
Elements.
Creation within any artistic genre or scientific paradigm
depends on constraints – the constraints of the genre, the
constraints of scientific data (Johnson-Laird, 1987). Artists
commonly invent new constraints when those of the
previous tradition no longer seem to be viable to them.
Indeed, creation depends on constraints, and the NONCE
hypothesis postulates that the greater the number of
constraints, within reason, the more creative individuals are
likely to be.
528
In order to test this conjecture, we asked participants to
generate creative sentences, and we manipulated the
constraints on the task in three independent ways. First, we
presented entities that had to be referred to in the sentences
either as line drawings of the entities, or as unambiguous
nouns referring to the entities. Second, the number of given
items to be incorporated in the sentence was either two or
three. In both cases, the participants had to include at least
three nouns in the sentence, but where they were given only
two nouns or two pictures, they were free to choose the
remaining noun. Third, the set of given entities either
included one animate entity or was entirely inanimate.
The pictures in our experiments were more constraining
than their corresponding nouns. A picture of an entity such
as a banana shows a particular instance of the entity with a
particular shape seen from a particular point of view. It
presents much more information than the corresponding
noun given that the noun can be depicted in infinitely many
different ways. Hence, the pictures provided more
constraints than the nouns. The claim is likely to be true in
general provided that the pictures are unambiguous and can
be immediately identified as an instance of the same entity
that the noun denotes.
Three given items are obviously more constraining than
two. The presence of an animate entity should also be more
constraining than its absence. Animacy calls for action, and
the need to formulate an action that the agent can carry out
on entities selected at random from our pool of entities
should force individuals to be more creative than they would
otherwise.
The use of pictures as a constraint has a corollary. A
picture brings to mind visual properties of an entity more
often than the corresponding noun brings them to mind. The
features that are brought to mind, whether for words or
pictures, should also be influenced by the need to frame a
sentence. Sentences in English call for finite verbs, and in
the present task verbs need to relate noun phrases referring
to the given entities. Visual properties, however, tend to be
expressed in adjectives, and so pictures may provide a
further constraint on the task of generating sentences in our
task: they may make the task harder and thereby also
enhance creativity.
The NONCE hypothesis therefore predicts that people
should produce more creative sentences from pictures than
from words, from three items than from two, from sets of
Experiment 1
Method
Participants Twenty-two native speakers of English, who
were undergraduates at Princeton University, participated in
this study for course credit.
Design The participants were their own controls and
generated sentences in eight conditions depending on
whether the entities were presented as pictures or as nouns,
whether there were three or two given entities, and on
whether or not the entities included an animate agent. Each
participant carried out four trials in each condition, i.e., a
total of thirty-two trials, which were presented in different
random order. The contents of the entities were
counterbalanced across the participants so that they occurred
equally often as pictures and as nouns in the experiment as a
whole.
Materials The pictures were line drawings selected from
Snodgrass and Vanderwart’s (1980) set of 260 schematic
pictures of concepts, most of which included exemplars
from Battig and Montague’s (1969) category norms and all
denoted basic level concepts (Rosch, Mervis, Gray,
Johnson, & Boyes-Braem, 1976). We selected only
unambiguous pictures, and excluded pictures of parts of
objects and pictures that elicited polysemous names (such as
saw and iron). The nouns were the names of the entities in
the line drawings and were presented in capital letters. None
of the entities in a set belonged to the same semantic
category.
Procedure The participants were instructed to generate a
creative sentence for each of the sets of entities. The
sentence had to include all of the given entities, which were
presented simultaneously on the screen of a computer, and,
whenever a set contained two entities and a question mark,
participants had to incorporate in the sentence a third noun,
which they were free to choose. No instructions elaborated
on the meaning of “creative”, and so the participants used
their own judgment on the matter. Four practice trials
occurred at the start of the experiment to familiarize the
participants with the task. They were told that when a set of
entities appeared on the screen, they should begin to think of
a sentence that referred to them. As soon as they had such a
sentence in mind, they had to press the space bar on the
computer keyboard, which removed the entities from the
screen. They then wrote down the sentence verbatim. The
next trial began only when the participant pressed the space
bar. Thus, for each trial, the computer recorded two
response times: the generation time, which was the time a
participant took to think of the sentence while the entities
were on the screen, and the writing time, which was the time
the participant took to write the sentence down while the
screen was blank.
529
Results
Creativity We used two measures to assess creativity. First,
on the assumption that creative sentences should be less
predictable, we used Shannon’s statistical measure of
information (Shannon & Weaver, 1949) to assess the
unpredictability of the sentences in the different conditions.
Second, we used a panel of judges to rate the creativity of
the sentences.
The sentences were sorted blindly into categories based
on the similarity of their meanings. For example, for the set
of entities LION STRAWBERRY HARP, sentences such
as: “After the lion finished playing the harp, he ate some
strawberry” and “The lion was playing the harp while eating
the strawberry” were put in the same category, whereas:
“The harp had a strawberry-colored lion carved in its post”
was put in a different category. The information-theoretic
measure, Â(-p log2 p), was computed for each of the eight
conditions. Figure 1 shows the means for each condition.
The sentences created from pictures were significantly more
informative than those created from the words (Wilcoxon
test, z = 3.01, p < .003). None of the other variables
produced reliable effects and there were no reliable
interactions among the variables.
Amount of information
entities including an animate agent than from inanimate sets.
Experiment 1 was designed to test these predictions.
3.5
3.3
3.1
2.9
2.7
2.5
2.3
2.1
1.9
1.7
1.5
Words
Pictures
2Anim.
2Inanim.
3Anim.
3Inanim.
Number (2 or 3) and type (one animate
or all inanimate) of given entities in a set
Figure 1: Information-theoretic measure of creativity in
Experiment 1
Two independent judges rated blindly the creativity of
each sentence on a seven-point scale (in which 1 denoted a
sentence that was not at all creative and 7 denoted an
extremely creative sentence). The judges’ ratings were
reliably correlated (Pearson’s r = .43, p < .001). The
measure of information also correlated reliably with the
judges’ ratings (Pearson’s r = .43, p < .001). Figure 2 shows
the mean ratings for each condition. The sentences
generated from pictures had reliably higher ratings of
creativity than the sentences generated from words.
(Wilcoxon test, z = 2.24, p = .029). The ratings from one of
the judges indicated another reliable difference: the
sentences from three entities were considered more creative
than the sentences from two entities (Wilcoxon test, z =
2.95, p < .004). None of the other variables produced
Judges' ratings
reliable effects and there were no reliable interactions
among the variables.
4
3.8
3.6
3.4
3.2
3
2.8
2.6
2.4
2.2
2
Words
Pictures
3Anim.
2Anim.
3Inanim.
2Inanim.
Number (2 or 3) and type (one animate
or all inanimate) of given entities in a set
had to refer to three entities, and so the experiment may not
have been powerful enough to detect any effect on creativity
of whether the third entity was given to the participants or
chosen by them. Likewise, the inclusion of an animate agent
in the set of entities yielded only one reliable effect: the
resulting sentences were shorter than those that were not
required to refer to an animate entity. The presence of an
animate entity in a set (e.g., COW BICYCLE ?) often led
participants to anthropomorphize it (e.g., “The cow rode the
bike to get milk from the corner grocery store”). This
tendency may explain why the resulting sentences were
more parsimonious. Yet, it is puzzling that it did not yield
more creative sentences.
Experiment 2
Figure 2: Creativity ratings in Experiment 1
Generation times We excluded the generation times for
three participants because they often responded before they
had thought of a sentence. Some of the remaining
participants occasionally made the same error and these
trials were also excluded from the analysis. The participants
took on average over two seconds longer to create a
sentence from line drawings (20.64 secs.) than from nouns
(18.41 secs.; Wilcoxon test, z = 2.21, p = .027). They also
took almost 4.5 seconds longer to create sentences about
three given entities (21.75 secs.) than about two (17.29
secs.; Wilcoxon test, z = 3.09, p = .002). No other
significant main effects or interactions occurred.
Sentence length The participants created longer sentences
from pictures (14.15 words) than from words (12.79 words;
Wilcoxon test, z = 3.56, p < .001), longer sentences from
three entities (14.56 words) than from two (12.37 words;
Wilcoxon test, z = 4.08, p < .001), and longer sentences
from sets that did not include an animate agent (14.16
words) than from those that did (12.78 words; Wilcoxon
test, z = 2.99, p < .005). There was a significant correlation
between generation times and sentence length (Pearson’s r =
.333, p < .001).
Discussion
The participants generated sentences of a greater creativity
in response to pictures than to words, as shown both by the
judges’ ratings and the statistical measure of
unpredictability. The NONCE hypothesis explains this
result on the grounds that line drawings of entities are more
constraining than nouns denoting them. The participants
also took longer to generate sentences from pictures than
from words, and the sentences that they generated from
pictures were also longer than those that they generated
from words. However, the correlation between sentence
length and creativity was not reliable. The given number of
entities affected latency and length of sentence in a
predictable way: the participants took longer to create
longer sentences from three entities than from two. But, this
variable had no reliable effect on the creativity of the end
result. In both cases, of course, the participants’ sentences
530
The second experiment examined further the differences
between line drawings and nouns. First, it aimed to validate
the prediction that the difference applies only to the
production of creative sentences. Hence, it compared the
production of sentences that were intended to be creative
with those that were not intended to be creative. Second, it
examined the nature of the nouns that the participants chose
in the two items condition.
Method
Participants We tested thirty-two native speakers of
English from the same population as before.
Design, procedure, and materials The participants acted as
their own controls and carried out two blocks of trials,
which were counterbalanced in order. In one block, the
participants were told to produce creative sentences, and in
the other block they were told to produce the first sentence
that came to mind. Within each block, half the trials used
pictures and half the trials used words (from the same
materials as Experiment 1). Half of all the sets included an
animate entity and the other half did not. There were six
trials in each of these four conditions, making a total of 24
trials in each block. The trials were in a different random
order for each participant. There were two given entities on
each trial and a question mark. The participants had to refer
to three entities in each sentence, and the question mark
reminded them to choose a third entity. They were told to
underline this third noun when they wrote down the
sentence. As in the first study, the position of the animate
entity in the set was counterbalanced, i.e., it occurred either
first or second in the set. All other aspects of the materials
and procedure were identical to those of Experiment 1.
Results
Creativity Figure 3 presents the means of Shannon's
measure of information, and Figure 4 presents the means of
the two judges' ratings on the seven-point scale. These
ratings correlated reliably (Pearson’s r = .331, p = .001), and
they also correlated reliably with the measure of information
(Pearson’s r = .545, p < .001). The sentences in the creative
condition were more informative than those in the non-
Amount of information
creative condition (Wilcoxon test, z = 5.06, p < .001). A
similar reliable effect occurred in the judges' ratings
(Wilcoxon test, z = 5.30, p < .0001). The creative condition
yielded significant effects of the variables predicted to
influence creativity. In this condition, the pictures yielded
sentences that were more informative and rated as more
creative than those produced from the words (Wilcoxon
tests, z = 2.80, p = .005; and
z = 3.16,
p = .002,
respectively). The sentences from inanimate entities were
also were also more informative than the sentences from
sets including an animate entity (Wilcoxon test, z = 2.20, p
= .028). No other main effects or interactions were reliable.
3.5
3.3
3.1
2.9
2.7
2.5
2.3
2.1
1.9
1.7
1.5
General Discussion
Words
Pictures
CrAnim.
CrInanim.
NcrAnim.
NcrInanim.
Type of instructions (creative or
noncreative) and type (one animate or all
inanimate) of given entities in a set
Figure 3: Information-theoretic measure of creativity in
Experiment 2
Judges' ratings
(Wilcoxon test, z = 3.21, p = .001) and the non-creative
condition (Wilcoxon test, z = 2.82, p = .005). And they were
longer for sets of inanimate entities than for sets including
an animate entity for both the condition (Wilcoxon test, z =
2.41, p = .016, and z = 2.77, p = .006, respectively).
The participants always had to think of a third word of
their own to include in their sentences, and they underlined
this word when they wrote the sentences down. When they
created a sentence from pictures, they tended to choose a
third word from the same category as one of the two entities
in the pictures (on 57% of relevant trials). But, when they
created a sentence from nouns, they were less inclined to
choose a third word from the same category (50%). This
interaction was reliable (Wilcoxon test, z = 2.35, p = .019).
3.8
3.6
3.4
3.2
3
2.8
2.6
2.4
2.2
2
Words
Pictures
CrAnim.
CrInanim.
NcrAnim.
NcrInanim.
Type of instructions (creative or
noncreative) and type (one animate or all
inanimate) of given entities in a set
Figure 4: Creativity ratings in Experiment 2
Generation times Overall, the participants took longer to
produce creative sentences (15.30 secs) than those that first
came to mind (9.46 secs; Wilcoxon test, z = 6.00, p <
.0001). They also took longer to produce the first sentence
that came to mind given pictures (10.13 secs) than given
words (8.68 secs; Wilcoxon test, z = 2.63, p < .01).
Sentence length The sentences were longer in the creative
condition (13.85 words) than in the noncreative condition
(11.14 words; Wilcoxon test, z = 8.32, p < .0001). They
were also longer in response to line drawings (14.17 words)
than to nouns (13.53 words), in both the creative condition
531
The goal of the present study was to test the hypothesis that
constraints enhance creativity. Line drawings of entities are
more constrained than nouns denoting those entities.
Pictures show a particular entity; words are general. That is
why, as the saying goes, "One picture is worth a thousand
words". In our first experiment, individuals had to produce
creative sentences referring to given entities. They were
more successful from entities depicted in line drawing than
from entities named by nouns. That is, their sentences were
judged to be more creative, and their sentences were less
predictable according to Shannon's information-theoretic
measure. An alternative explanation of the phenomenon is
that individuals find it harder to put pictures into sentences
than to put words into sentences. This account is consistent
with the findings of our first experiment: the participants did
indeed take longer to frame sentences based on pictures than
sentences based on words, and the sentences based on
pictures contained more words than the sentences based on
words. However, one result that goes against this account is
that the number of entities to which the sentences had to
refer yielded reliable effects on both latency and length, but
did not show any such effect on our measures of creativity.
It may be that the experiment was not sensitive enough to
detect such an effect, which the hypothesis concerning
constraints, certainly predicts. But, fortuitously, the lack of
this effect, in contrast to those on latency and length, counts
against the notion that creativity is merely an effect of how
much time individuals devote to constructing sentences.
Our second experiment revealed an interaction that is
crucial for the hypothesis about constraints. The experiment
demonstrated the critical role of the instructions to generate
a creative sentence. It contrasted what happened in this
condition with a condition in which the participants merely
stated the first sentence that came to mind. Pictures yielded
sentences that were more imaginative than the sentences
generated from words. But, this effect occurred only when
the participants had been told to produce a creative
sentence. Once again, the results ruled out the simple
notion that creativity depends solely on the time allotted to
the task.
Our task poses problems to participants. They have to
solve the problem of formulating a sentence that makes
reference to certain entities. Hence, they have a goal and
they have constraints on how to reach that goal. The
constraints depend on whether the set of entities is presented
pictorially or verbally. The particular sentences that the
participants created provide us with clues about the nature
of the problem-solving process. Pictures, as we suggested in
the Introduction, make the visual properties of objects
salient, and these properties often go on to guide the
construction of sentences. For instance, given pictures of the
following entities: LAMP PEAR NECKLACE, a participant
created the sentence: “Although she was shaped like a pear,
I gave her a necklace with the lamp light on.” Similarly,
given pictures of the entities: LOBSTER FENCE
UMBRELLA, another participant wrote: “Beyond those
white picket fences, the world is a rainy place, tough as a
lobster’s exoskeleton, and you’ve got no umbrella.” And, in
Experiment 2, there was a common tendency to respond to:
ONION BOTTLE ? by creating a sentence in which the
onion was in the bottle (cf. Tabossi, Colombo, & Jobs,
1987).
In general, individuals are likely to retrieve certain
knowledge about the entities in question, depending in part
on their format of presentation, and then to use this
knowledge as a constraint on the creation of a scenario that
can be described in a sentence referring to the entities. One
sign of this process was illustrated by the inclusion of an
animate agent in the set of entities (see Bock, Loebell, &
Morey, 1992; McDonald, Bock, & Kelly, 1993). There was
then an overwhelming tendency (in Experiment 2) for
individuals to construct a sentence referring to an action
carried out by the animate agent on the inanimate entity.
This tendency, almost automatic, yielded shorter and more
stereotyped sentences.
After the automatic activation of properties, individuals
search for an appropriate scenario. Some properties may
suffice for the creation of a relation between the entities.
When the participants had to generate the first sentence that
came to mind, they would sometimes merely enumerate the
existence of the relevant entities in some setting, e.g., “Mary
found a pear, a necklace, and a lamp in her closet.” But, the
participants tended to avoid such lists when they had to be
creative.
The need for "freedom to create" is commonplace. But,
according to the NONCE hypothesis about creativity, such
freedom is an illusion. On the contrary, constraints are at the
heart of the creative process. They govern the generation
of ideas, and they provide criteria for the evaluation of
ideas. Without constraints, there is no creativity. Hence, the
creators of works of art often go out of their way to seek
constraints, e.g., the serial method of musical composition
developed by Schönberg, the invention of highly
constraining verse forms such as the villanelle and sestina,
and the elaborate formal and combinatorial constraints
devised by OULIPO (the Ouvroir de Littérature Potentielle,
the European group of writers set up by Raymond Quenneau
532
and others). The experiments reported here corroborate this
account at least in a preliminary way. When individuals
create sentences to refer to given entities, the constraints
embodied in the representation of these entities can help
them to be more creative. Pictures are a better source of
creativity than words.
Acknowledgements
We thank Sam Glucksberg, Geoff Goodwin, Uri Hasson,
Louis Lee, and Sanna Reynolds for their helpful comments.
References
Battig, W. F., & Montague, W. E. (1969). Category norms
for verbal items in 56 categories: A replication and
extension of the Connecticut category norms. Journal of
Experimental Psychology Monograph, 80, 1-46.
Bock, J. K., Loebell, H., & Morey, R. (1992). From
conceptual roles to structural relations: Bridging the
syntactic cleft. Psychological Review, 99, 150-171.
Chomsky, N. (1965). Aspects of The Theory of Syntax.
Cambridge, MA: MIT Press.
Finke, R. A., Ward T. B., & Smith, S. M. (1992). Creative
Cognition: Theory, Research, and Applications.
Cambridge, MA: MIT Press.
Johnson-Laird, P. N. (1987). Reasoning, imagining, and
creating. Bulletin of the British Psychological Society, 40,
121-129.
Johnson-Laird, P. N. (1993). Human and Machine Thinking.
Hillsdale, NJ: Lawrence Erlbaum Associates, Publishers.
Johnson-Laird, P. N. (2002). How jazz musicians improvise.
Music Perception, 19, 415-442.
Levelt, W. J. M. (1989). Speaking: From Intention to
Articulation. Cambridge, Massachussetts: MIT Press.
McDonald, J. L., Bock, K., & Kelly, M. H. (1993). Word
and world order: Semantic, phonological, and metrical
determinants of serial position. Cognitive Psychology, 25,
188-230.
Rosch, E., Mervis, C. B., Gray, W. D., Johnson, D. M., &
Boyes-Braem, P. (1976). Basic objects in natural
categories. Cognitive Psychology, 8, 382-439.
Shannon, C. E., & Weaver, W. (1949). The Mathematical
Theory of Communication. Urbana: University of Illinois
Press.
Snodgrass, J. G., & Vanderwart, M. (1980). A standardized
set of 260 pictures: Norms for name agreement, image
agreement, familiarity, and visual complexity. Journal of
Experimental Psychology: Human Learning and Memory,
6, 174-215.
Tabossi, P., Colombo, L., & Jobs, R. (1987). Accessing
lexical ambiguity: Effects of context and dominance.
Psychological Research, 49, 161-167.